Shared memory for your team's agents
Persistent memory that survives across sessions and detects when agents contradict each other.
There are a lot of tools that give agents memory. Engram works alongside any of them.
What makes Engram different: it catches when agents contradict each other. One agent learns something, another agent learns something else, and now they disagree. Engram detects that and surfaces it before it becomes a bug.
Your whole team shares the same memory. Your agents, your teammates' agents — everyone stays in sync. The same context, the same facts, no duplicated mistakes.
Right now it's your personal memory layer that survives across sessions. The bigger picture is a consistency layer for teams of hundreds of agents working on the same codebase at once.
Step 1 — Create an account
Sign up at engram-memory.com/dashboard. Create a new workspace or join an existing one using an invite key from a teammate.
Step 2 — Run the installer
macOS / Linux:
curl -fsSL https://engram-memory.com/install | shWindows PowerShell:
irm https://engram-memory.com/install.ps1 | iexWindows CMD:
curl -fsSL https://engram-memory.com/install.cmd -o install.cmd && install.cmd && del install.cmdStep 3 — Restart your editor, then ask your agent:
"Set up Engram for my team"
Your agent handles the rest.
Engram is currently optimized for MCP-native workflows in:
Each guide includes the expected MCP config path, restart step, verification flow, and common setup mistakes.
If you want to run Engram from this repository during development:
pip install -e ".[dev]"
python -m engram.cli serve --httpThen open:
http://127.0.0.1:7474/dashboard
If engram is not on your PATH, python -m engram.cli ... works reliably.
Create an account at engram-memory.com to start a workspace. A demo video is coming soon that will walk through the full setup flow.
Your data is encrypted in transit and at rest, fully isolated per workspace, and never read, analyzed, trained on, or shared with anyone. Delete your workspace and everything is gone.
| Tool | Purpose |
|---|---|
engram_commit |
Persist a verified discovery |
engram_query |
Pull what your team's agents know |
engram_conflicts |
Surface contradictions |
engram_resolve |
Settle disagreements |
engram_promote |
Graduate ephemeral memory to durable |
engram install # Auto-detect IDEs and configure MCP
engram serve # Start MCP server (stdio mode)
engram serve --http # Start MCP server (HTTP mode)
engram setup # One-command workspace setup
engram status # Show workspace status
engram info # Display detailed workspace info
engram whoami # Show current user identity
engram search <query> # Query workspace from terminal
engram stats # Show workspace statistics
engram config show # Display configuration
engram config set <key> # Update configuration
engram tail # Live stream of workspace commits
engram verify # Verify installation
engram doctor # Diagnose setup issues
engram completion <shell> # Install shell tab completionRuns asynchronously in the background:
| Tier | Method | Catches |
|---|---|---|
| 0 | Entity matching | "rate limit is 1000" vs "rate limit is 2000" |
| 1 | NLI cross-encoder | Semantic contradictions |
| 2 | Numeric rules | Different values for same entity |
| 3 | LLM escalation | Ambiguous cases (rare, optional) |
Commits return instantly. Detection completes in the background (~2-10s on CPU).
Not everything deserves to stick around. Scratchpad facts expire in 24h, unverified observations after 90 days. Decisions and confirmed facts are kept forever. Old context stops crowding out what matters now.
Engram exists because of a paper.
Multi-Agent Memory from a Computer Architecture Perspective: Visions and Challenges Ahead — Yu et al. (2026), UCSD SysEvol — is the primary intellectual foundation of this project. It reframes multi-agent memory as a computer architecture problem: coherence, consistency, and shared state across concurrent agents. That framing is what Engram is built to implement in practice.
The rest of the literature informs specific subsystems:
- Xu et al. (2025) — A-Mem's Zettelkasten structure for fact enrichment
- Rasmussen et al. (2025) — Graphiti's bitemporal modeling for temporal validity
- Hu et al. (2026) — Survey confirming shared memory as an open frontier
- Alqithami (2025) — FiFA: forgetting-by-design improves agent coherence
Full literature review: docs/LITERATURE.md
Implementation details: docs/IMPLEMENTATION.md
PRs welcome. See CONTRIBUTING.md.
For a full description of the test suite — what each module covers and the per-test breakdown for lifecycle and conflict tests — see tests/TESTS.md.